A Phenotypic Extraction and Deep Learning-Based Method for Grading the Seedling Quality of Maize in a Cold Region

Abstract

Background: Low-temperature stress significantly restricts maize germination, seedling growth and development, and yield formation. However, traditional methods of evaluating maize seedling quality are inefficient. This study established a method of grading maize seedling quality based on phenotypic extraction and deep learning. Methods: A pot experiment was conducted using different low-temperature combinations and treatment durations at six different stages between the sowing and seedling phases. Changes in 27 seedling quality indices, including plant morphology and photosynthetic performance, were investigated 35 d after sowing and seedling quality grades were classified based on maize yield at maturity. The 27 quality indices were extracted, and a total of 3623 sample datasets were obtained and grouped into training and test sets in a 3:1 ratio. A convolutional neural network-based grading method was constructed using a deep learning model. Results: The model achieved an average precision of 98.575%, with a recall and F1-Score of 98.7% and 98.625%, respectively. Compared with the traditional partial least squares and back propagation neural network, the model improved recognition accuracy by 8.1% and 4.19%, respectively. Conclusions: This study provided an accurate grading of maize seedling quality as a reference basis for the standardized production management of maize in cold regions.

1. Introduction

Northeast China, an important commercial grain production base with approximately 6,830,700 ha dedicated to spring maize cultivation, accounting for more than 30% of the nation’s maize area and 32.8% of its maize production, is important for China’s agricultural production and food security [1]. Maize growth and development are very vulnerable to low temperatures. Maize requires warmth and rainfall during its growing season, and Northeast China, especially its high-latitude cold regions, experiences significant year-to-year fluctuations in temperature. On average, the region’s maize production experiences chilling injury every 3 to 4 years. Severe events in 1969, 1972, and 1976 resulted in an average crop yield reduction of 5.78 billion kg in the Northeast region [2,3]. Similarly, while the increased temperature due to global warming is generally favorable for maize production in Northeast China, this is still concentrated in the winter months, whereas insignificant warming occurs during the maize growing season [4,5,6,7]. Consequently, severe low-temperature chilling injury is more frequent in Northeastern China, especially in Heilongjiang and Jilin provinces, and its incidence may increase significantly in the early (from sowing to seven leaves) and late (from milky ripening to ripening) growth stages [8,9,10].

The region’s spring maize producers use direct sowing technology and reasonably dense planting [11], timely and early sowing [12], no-tillage sowing [13], and other cultivation techniques. However, the slow warming of the tillage soil and stage of low temperature after sowing often lead to seed germination and seedling formation challenges, including reduced and delayed seedling emergence, decreased seedling rate, and an increased proportion of weak seedlings. In severe cases, seeds completely lose their viability, resulting in a large area without seedlings [14,15,16]. Ultimately, these factors make it difficult to establish high-quality maize populations in the region, causing a major bottleneck for maize mechanization and efficient production. Consequently, the region’s maize industry’s efforts to increase efficiency and achieve sustainable development are hindered.

Previous research groups have conducted extensive studies on the response patterns of maize to low temperatures from sowing to seedling. Maize is more sensitive to low temperature in its early developmental stage (seed germination, seedling emergence, and vegetative stages) than in the latter developmental stages. Low-temperature stress affects the content of hormones, such as abscisic acid, cytokinin, and gibberellins, in seedlings and inhibits antioxidant enzyme activities and protein synthesis [17,18,19]. In addition, it reduces seed germination rate, germination potential, seed viability, and seedling emergence, which are detrimental to seedling morphology and photosynthetic performance, leading to a decline in seedling quality, a shortening of the grouting period, an increase in the rate of empty culm and bald tips, and insufficient panicles and grains, ultimately limiting grain yield [20,21,22]. Verheul et al. [23] showed that each 0.7 °C drop in temperature during maize growth and development delayed maturity by 7 d and resulted in an 8% reduction in yield. Fryer et al. [24] found that as the temperature of the growing environment decreased, maize leaf growth slowed, and the plant’s net assimilation rate, relative leaf area, and relative growth rate were decreased [25]. Furthermore, low-temperature stress resulted in reduced primary root length and lateral root number in maize [26,27,28]. Sub-optimal temperature conditions also affected leaf and root extension, root area to aboveground area ratio, and root length to leaf area ratio of maize seedlings, consequently affecting grain yield [29,30]. Thus, achieving excellent seedling quality is an important way to achieve high yields of intensively planted maize in this region. Therefore, based on the quality of maize seedlings in the field, accurate assessment of the potential impact of unfavorable meteorological conditions on maize growth, development, and yield from sowing to the seedling stage, and efficient cultivation and control techniques to exploit the advantages and avoid disadvantages are significant for fully exploiting the potential of maize for increased yields through dense planting and efficient utilization of regional agricultural resources. Most traditional methods for predicting and evaluating the quality of maize seedlings are based on experts’ sensory perceptions and chemical index measurements [21,31], which are time and energy consuming and rely on prior farming experience. Consequently, employing scientific, accurate, and rapid detection technologies for evaluating the quality of maize seedlings in the field is important for promoting efficient maize production.

In the wake of deep learning, scholars have increasingly applied convolutional neural networks (CNNs) in agronomy research, with relatively good results in the identification of pests and diseases [32], classification and screening of weeds in the field [33], collection of key information about a crop [34], and crop growth prediction [35]. For instance, using a fine-tuned Mask region-based CNN (R-CNN) model, Gao et al. [36] proposed an automatic maize seedling identification method that is adaptable to various developmental stages of maize seedlings to quickly and accurately extract phenotypic information from field environments while reducing labor costs. The model had an average detection accuracy of 88.70%, with average accuracies of detection of seedling emergence in 2019, 2020, and 2021 of 98.87%, 95.70%, and 98.77%, respectively. Similarly, a new automatic identification algorithm for multiple developmental stages of rice spikes (MDSRS) based on improved Faster R-CNN was proposed by Zhang et al. [37]. Rice spikes were accurately identified by acquiring image sequences of different developmental stages using ground-based monitoring equipment. The group showed that the model had an average accuracy of 92.47% in detecting the developmental period of rice spikes, with a 0.7–1.1 d error in identifying the onset dates for rice heading, milk maturity, and full maturity stages. Li et al. [34] presented a deep-learning-based canopy phenotyping method for classifying, segmenting, and counting the canopy leaves of watermelon seedlings using the Mask R-CNN network. They showed that the model had an average relative error of 2.33% for number of true leaves, 4.59% for number of cotyledons, 8.37% for leaf area, and 3.27% for plant height. Domestic and international research on seedling quality identification are mainly based on canopy image phenotypic indicators for single organ detection without involving the comprehensive monitoring and grading evaluation of maize seedling quality indicators. However, a deep learning-based evaluation model for multimodal maize seedling quality has rarely been reported.

In this study, the main cultivar of maize, ‘Xianyu335′, which is widely planted in the high-latitude cold region of Northeast China, was selected as the test material. An extraction technology for the seedling quality index was combined with a CNN to construct a method for grading cold region maize seedlings based on the extraction of seedling quality phenotypes and deep learning, thus unveiling a rapid and accurate grading of the quality of maize seedlings in cold regions. This provided the theoretical basis for the standardized and refined production management of maize throughout the cultivation process and a reference for the timely implementation of maize stress control techniques in cold regions, enabling farmers to maximize benefits while minimizing potential harm.

2. Materials and Methods

2.1. Experimental Design

2.1.1. Description of the Test Site

This study was conducted from 2019 to 2021 at the experimental site of Heilongjiang Bayi Agricultural Reclamation University (Heilongjiang, Daqing, 46°35′ N, 125°9′ E, 146 m.a.s.l). The region is in the north temperate continental monsoon climate zone. The average daily temperature and rainfall conditions during the 2019 to 2021 maize growing season are shown in Figure 1. For the potting test, 0–20 cm of plough layer farmland soil was used. The soil type was meadow soil, with the following basic physicochemical properties: alkaline dissolved nitrogen 104.46 mg·kg⁻¹, quick-acting phosphorus 11.47 mg·kg⁻¹, quick-acting potassium 113.73 mg·kg⁻¹, organic matter 26.59 g·kg⁻¹, and pH 8.06. The chemical properties of the irrigation water were as follows: electrical conductivity 0.48 dS·m⁻¹, pH 7.5, NO₃⁻ content 8 mg·L⁻¹, and sodium adsorption ratio 0.14, which are features of high-quality irrigation water for maize production [38].

2.1.2. Test Material Treatment

The main cultivar of maize, Xianyu335, widely planted in high-latitude cold regions of Northeast China, was used as the test material. Before the experiment, polyethylene pots (40 cm in diameter × 40 cm in height) were filled with air-dried, broken, and sieved soil. On 25th April, full-grown, uniformly sized maize seeds were selected, and 10 seeds were evenly sown in each pot at a depth of 4 cm. After sowing, the seeds were uniformly irrigated to ensure that 75% of the field water holding capacity of the soil in each pot was reached, followed by low-temperature stress treatment. The experiment was conducted in a 3-factor, completely randomized design. The main treatment was the low-temperature timed treatment, with six stages set: 1–5 (P1), 6–10 (P2), 11–15 (P3), 16–20 (P4), 21–25 (P5), and 26–30 (P6) d after sowing. The secondary treatment was the low-temperature treatment, with six stages set: 0, 2, 4, 6, 8, and 10 °C. The sub-treatment was the low-temperature duration treatment, with six stages set: 0 (D0), 1 (D1), 2 (D2), 3 (D3), 4 (D4), and 5 (D5). Among them, the low-temperature treatment 0 (D0) d is the control group, with a total of 180 treatment combinations and a control group (D0) without low-temperature treatment. Each treatment combination had 10 replicates for subsequent experiments, with each pot being used as a replicate each time.

Based on the treatment combinations, the corresponding polyethylene pots were placed in an artificial climate chamber for cultivation. During the low-temperature stress treatment and in addition to the treatment temperature, the humidity of the artificial climate chamber was set at (50 ± 5)%, the light duration at 14.5 h·d⁻¹, and the light intensity at 450 µmolm⁻²·s⁻¹. At the end of each treatment, the polyethylene pots were placed in the natural field environment for further cultivation. At 35 d after sowing, 4 pots were randomly selected from each surviving treatment combination, and 24 control treatment pots were selected (as control treatment was set for each low-temperature timed treatment, i.e., 6 × 4 = 24 pots). Two representative maize seedlings were selected from each pot to determine various quality indicators of seedlings. Theoretically, 4464 maize seedlings (180 × 4 × 2 × 3 + 24 × 2 × 3) were expected to be collected in the three-year experiment. However, some seedlings under the 2 °C low-temperature treatment could not survive, particularly in stages with 0 °C. Therefore, phenotypic agronomic traits of 3623 maize seedlings were collected over the three study years. Finally, the remaining six pots under each treatment combination and control treatment were intercropped, with one seedling left in each pot. The individual yield of maize in each treatment combination was investigated during the mature period. In the soil of each pot, 2.83 g of pure N, 1.47 g of P₂O₅, and 1.13 g of K₂O were applied, of which 30% of pure N and all phosphorus and potassium fertilizers were applied as basal fertilizers, whereas the remaining 70% of pure N was applied as topdressing fertilizer at the V6 stage; the rest management was similar to that during field production.

2.1.3. Design of Individual Quality Indicators

Based on previous studies on crop seedling quality in the field [39,40,41,42] combined with the field growth characteristics of maize, and following the principles of dominance, practicability, and operability, the following indicators of seedling quality were used: plant height (x₁), stem base diameter (x₂), coleoptile length (x₃), leaf sheath length of first spreading leaf (x₄), leaf sheath length of second spreading leaf (x₅), length of first spreading leaf blade (x₆), width of first spreading leaf blade (x₇), length of second spreading leaf blade (x₈), width of second spreading leaf blade (x₉), length of third spreading leaf blade (x₁₀), width of third spreading leaf blade (x₁₁), total leaf area of first to third spreading leaves (x₁₂), primordial radicle length (x₁₃), primordial radicle diameter (x₁₄), number of secondary radicles (x₁₅), number of nodal roots (x₁₆), root volume (x₁₇), leaf relative chlorophyll content SPAD (x₁₈), initial fluorescence F₀ (x₁₉), maximal fluorescence F_m (x₂₀), maximum photochemical efficiency of PSII F_v/F_m (x₂₁), aboveground fresh weight (x₂₂), belowground fresh weight (x₂₃), residual seed fresh weight (x₂₄), aboveground dry weight (x₂₅), belowground dry weight (x₂₆), and residual seed dry weight (x₂₇), totaling 27 metrics.

2.2. Measurements

2.2.1. Determination of Morphological Indicators in Seedlings

First, distilled water was used to slowly rinse the soil at the roots of the collected intact corn seedlings. Absorbent paper was used to absorb surface moisture, and the various quality indicators of the seedlings were then measured.

(1)

Plant height (x₁): a ruler was used to measure the distance from the base of the stem (the junction between the stem and root) to the top of the longest leaf when upright;

(2)

Stem base diameter (x₂): the diameter of the stem base (the junction between the stem and root) was measured using a vernier caliper;

(3)

Coleoptile length (x₃): a ruler was used to measure the length of the conical sheath outside the embryo;

(4)

Leaf sheath length of first spreading leaf (x₄): the length of the base of the first fully unfolded leaf that surrounds the stem in a sheath shape was measured using a ruler;

(5)

Leaf sheath length of second spreading leaf (x₅): the length of the base of the second fully unfolded leaf that surrounds the stem in a sheath shape was measured using a ruler;

(6)

Length of first spreading leaf blade (x₆): the length from the base to the tip of the first fully unfolded leaf blade was measured using a ruler;

(7)

Width of first spreading leaf blade (x₇): the width at the widest point of the first fully unfolded leaf blade was measured using a ruler;

(8)

Length of second spreading leaf blade (x₈): the length from the base to the tip of the second fully unfolded leaf blade was measured using a ruler;

(9)

Width of second spreading leaf blade (x₉): the width at the widest point of the second fully unfolded leaf blade was measured using a ruler;

(10)

Length of third spreading leaf blade (x₁₀): the length from the base to the tip of the third fully unfolded leaf blade was measured using a ruler;

(11)

Width of third spreading leaf blade (x₁₁): the width at the widest point of the third fully unfolded leaf blade was measured using a ruler;

(12)

Total leaf area of first to third spreading leaves (x₁₂): the length width coefficient method (length × width × 0.75) was used to calculate the single leaf area of the first to third fully unfolded leaves, and then the total leaf area was calculated [43].

(13)

Primordial radicle (the main root that grows first from the seed embryo) length (x₁₃), diameter (x₁₄), number of secondary radicles (young roots gradually growing on both sides of the primary embryonic root, x₁₅), number of nodal roots (roots growing on stem nodes, x₁₆), and root volume (the size of the space occupied by the root system, x₁₇) were determined using the WinRHIZO Pro (Zealquest Scientific Technology Co., Ltd., Shanghai, China) root analysis system [44].

2.2.2. Detection of Chlorophyll Content and Chlorophyll Fluorescence Parameters in Leaves

Leaf relative chlorophyll content SPAD (x₁₈), initial fluorescence F₀ (x₁₉), maximal fluorescence F_m (x₂₀), and maximum photochemical efficiency of PSII F_v/F_m (x₂₁) were estimated. The latest fully expanded functional leaves of maize seedlings were taken, and following dark adaptation for 30 min, the chlorophyll fluorescence parameters of the upper, middle, and lower three points of leaves were determined using the OS-30p chlorophyll pulse-made fluorescence analyzer (Opti-Sciences, Inc., Hudson, NY, USA). The initial fluorescence F₀ was read, and the maximum fluorescence F_m was obtained by irradiating with saturated pulsed light after the stabilization of F₀, and the maximum photochemical efficiency was calculated using the formula F_v/F_m = (F_m − F₀)/F_m. The relative chlorophyll content was determined at the same locations using the SPAD-502 Plus chlorophyll meter (Konica Minolta Holdings, Inc., Tokyo, Japan), and finally, the average value was taken [45,46].

2.2.3. Determination of Fresh and Dry Weight of Various Parts of Seedlings

The plant was divided into three parts: aboveground, belowground, and residual seeds (seed-related tissue parts remaining after germination). The aboveground fresh weight (x₂₂), belowground fresh weight (x₂₃), and residual seed fresh weight (x₂₄) were determined, and the samples were then put in a kraft paper bag to at 105 °C for 30 min to halt all cell activity. Subsequently, the samples were dried at 80 °C until constant weight. The aboveground dry weight (x₂₅), belowground dry weight (x₂₆), and residual seed dry weight (x₂₇) were determined.

2.2.4. Measurement and Calculation of Yield Per Plant

All maize plants for each treatment combination were harvested manually at the maturity stage to determine kernel yield. After measuring the kernel weights from five randomly selected plants from each treatment, kernel moisture content was determined using the PM-8188-A Grain Moisture Meter (Kett Electric Laboratory Co., Ltd., Tokyo, Japan), and the final grain yield was standardized to 14.0% moisture content [47].

2.3. Constructing a Grading Model for the Quality of Cold-Land Maize Seedlings

As part of the field of artificial intelligence, deep learning has recently emerged as a new research direction and has already achieved breakthroughs in multiple categories of applications such as natural language processing, pattern recognition, and many applications in recognition and classification capabilities in multiple domains [48]. According to the link between maize seedling quality indices and single plant yield, we constructed a cold-land maize seedling quality grading model based on maize phenotypic parameters using CNN, a classical deep learning model.

2.3.1. CNN Model Building

CNN is a deep feedforward neural network with local connections and weight sharing that differs from other neural networks such as recurrent and depth neural networks. Its main feature is convolution operation [49]. CNN comprises a convolutional layer, an activation function layer, a pooling layer, and a fully connected layer that are cross-stacked. The common CNN is a two-dimensional network structure that typically conducts feature extraction from pictures. As the sample data in this study were one-dimensional maize seedling quality indicators, a one-dimensional CNN model (LeNet-5) was used for maize seedling quality grading (Figure 2).

The calculation process for each layer in Figure 2 was as follows:

Convolutional layer: Feature extraction was performed using the convolutional layer, with more features being extracted with more convolutional kernels. The sample data were convoluted from the input layer to the C1 convolution layer, which comprised a series of convolution kernels that could be obtained through learning, a core component of CNN. The function of the convolution layer is to extract features and compress the amount of data. In this study, we set ‘padding = same’ to ensure that the output eigenmatrix of the convolution layer was the same size as that of the input eigenmatrix. C1, C3, and C5 had convolution kernel sizes of 64, 32, and 32, and convolution window sizes of 5, 7, and 6, respectively. The convolution operation was performed based on the size of the convolution kernel and convolution window. The convolution operation formula was as follows:

$$y^j=f\left({\displaystyle \sum _i{k}^{ij}}\ast x^i+b^j\right)$$

(1)

where $y^j$ and $x^i$ are the $i$ input feature and $j$ the output feature, respectively; $k^{ij}$ denotes the convolution kernel used for the convolution of this layer, while $\ast$ represents the convolution operation and $b^j$ is the bias of the $j$ feature. The Relu activation function was used for the nonlinear operation as follows:

$$f\left(x\right)=\max \left(0,x\right)$$

(2)

Pooling layer: Following the convolution layer was the S2 pooling layer, which reduced the data dimension, performed secondary feature extraction, and subsampled the data using the principle of local correlation to reduce the amount of data and enhance the robustness of the neural network [50]. The calculation formula was as follows:

$$s_j^{l+1}=f\left({\alpha }_j^{l}down\left(y_j^l\right)+b_j^{l+1}\right)$$

(3)

where $down(\cdot )$ sampling denotes the subsampling function, ${\alpha }_j^l$ and $b_j^{l+1}$ denote the layer $j$ weighting coefficient and bias coefficient, respectively. After the pool layer, a dropout layer was added to prevent overfitting, and neurons with 0.1 random inactivation ratios were added to speed up the convergence of the model and improve generalization.

Fully connected layer: During the fully connected layer, every neuron was connected to that in the previous layer. In this layer, the softmax function as used to input the output results of the network to the full connection layer for classification, and the matrix operation was transformed into eigenvectors of length 4. Corresponding to the four seedling quality classifications used in this study, the output vectors were converted into the expression of the classification probability of each category using the softmax function [51], as follows:

$$soft \max =\frac{\exp \left(z_i\right)}{{\displaystyle {\sum }_j\exp \left(z_j\right)}}$$

(4)

where $i$ and $j$ are elements of $z$, and $z$ is the input vector, respectively.

2.3.2. Optimizer

The seedling quality evaluation model established in this study used the Adam optimizer to calculate the optimal network parameters. Based on first- and second-order moment estimation of the gradient, the optimizer dynamically adjusted the learning rate for each parameter [52], accelerating the convergence and improving the performance of the model. The update steps of the Adam algorithm were as follows:

(1)

The gradient $g_t$ was first computed for the current time step, initialized to be computed from $T=1$, and $t$ time iterations of training were performed, with the parameter values being updated for each computation:

$$g_t=f_t\left({\theta }_{t-1}\right)$$

(5)

(2)

The exponential moving average $m_t$ of the gradient was calculated, $m_0$ was initialized to 0, and the ${\beta }_1$ coefficient was used as the exponential decay rate to control weight allocation. Typically, a value close to 1 was taken, set to 0.9 in this study.

$$m_t={\beta }_1{m}_{t-1}+\left(1-{\beta }_1\right)g_t$$

(6)

(3)

$v_0$ was initialized to 0 and the exponential moving average $v_t$ of gradient squared was calculated, with ${\beta }_2$ being the exponential decay rate coefficient, and the gradient squared was controlled before setting it to 0.99 in this study.

$$v_t={\beta }_2{v}_{t-1}+\left(1-{\beta }_2\right)g_t^2$$

(7)

(4)

Initially initializing $m_0$ to 0 could lead to $m_t$ leaning towards 0, particularly at the beginning of training. To reduce the impact of bias on the initial training stage, bias correction was applied to the gradient mean $m_t$.

$${\widehat{m}}_t=\frac{m_t}{1-{\beta }_1^t}$$

(8)

(5)

Similar to $m_0$, deviation correction was performed on $v_t$.

$${\widehat{v}}_t=\frac{v_t}{1-{\beta }_2^t}$$

(9)

(6)

The parameters were updated, and the learning rate was set to $\alpha$ = 0.001 and $\epsilon$ = 10⁻⁸.

$${\theta }_t={\theta }_{t-1}-\alpha \frac{{\widehat{m}}_t}{\epsilon +\sqrt{{\widehat{v}}_t}}$$

(10)

2.3.3. Model Evaluation Indicators

In this study, the collected maize seedling quality data were inputted to the convolution neural network for training, and the accuracy, precision, recall rate, F1-Score and loss were selected as the model prediction effect evaluation indices.

The accuracy of a classification model is typically measured by dividing the number of samples correctly predicted by the model by the total number of samples. Models with higher accuracy perform better in classification. The formula for calculating accuracy was as follows:

$$accuracy=\frac{TP+TN}{TP+FP+FN+TN}$$

(11)

Precision rate was used to describe the proportion of all positive samples predicted by the model that were truly positive and was calculated as follows:

$$precision=\frac{TP}{TP+FP}$$

(12)

Recall indicated the proportion of samples that were predicted to be positive out of those that were actually positive and was calculated as follows:

$$recall=\frac{TP}{TP+FN}$$

(13)

A model’s F1-Score measures a model’s balance of predictive accuracy and coverage of positive class samples, which is the reconciled mean of precision and recall. The F1-Score was calculated as follows:

$$F1Score=\frac{2\times precision\times recall}{precision+recall}$$

(14)

In this study, TP (True Positive) was a positive sample predicted to be positive by the model, FP (False Positive) was a negative sample predicted to be positive by the model, FN (False Negative) was a positive sample predicted as negative by the model, and TN (True Negative) was a negative sample predicted as negative by the model.

3. Results

3.1. Data Pre-Processing and Analysis

3.1.1. Effect of Different Stages of Low-Temperature Treatment and Duration on Seedling Quality Indicators from Sowing to Seedling Stage

We collected data and plotted graphs using MS Excel 2010 (Microsoft Corporation, Redmond, WA, USA) and Origin 2018 software (OriginLab, Northampton, MA, USA). We plotted line graphs (Figure S1) for the 27 seedling quality indices of maize plants under each treatment combination. These graphs revealed an overall decreasing trend for 17 individual quality indices with the increasing duration of low-temperature treatment. These indices included overall plant height (x₁), stem base diameter (x₂), coleoptile length (x₃), leaf sheath length of first spreading leaf (x₄), leaf sheath length of second spreading leaf (x₅), length of first spreading leaf blade (x₆), width of first spreading leaf blade (x₇), length of second spreading leaf blade (x₈), width of second spreading leaf blade (x₉), length of third spreading leaf blade (x₁₀), width of third spreading leaf blade (x₁₁), total leaf area of first to third spreading leaves (x₁₂), root volume (x₁₇), aboveground fresh weight (x₂₂), belowground fresh weight (x₂₃), aboveground dry weight (x₂₅), and belowground dry weight (x₂₆). The chlorophyll fluorescence parameters F₀ (x₁₉) and F_m (x₂₀) showed the lowest values at 2 d of low-temperature treatment and were then stabilized at a lower level with prolonged low-temperature treatment; conversely, the residual seed fresh weight (x₂₄) and residual seed dry weight (x₂₇) of residual seeds showed an increasing trend with prolonged low-temperature treatment.

To further understand the effect of different low-temperature treatments on the 27 maize seedling quality indicators, we draw heat maps using Tbtools (v 1.120) software [53], and analyzed the trends of the 27 maize seedling quality indicators under low-temperature treatment. As shown in Figure 3, the 27 seedling quality indicators at different stages and under different combinations of low-temperature treatments generally showed three trends, with the diameter of the primordial radicle diameter (x₁₄), number of nodal roots (x₁₆), chlorophyll content (x₁₈), fluorescence parameters F₀ (x₁₉), F_m (x₂₀), and F_v/F_m (x₂₁), residual seed fresh weight (x₂₄), and residual seed dry weight (x₂₇) exhibiting a similar pattern. We observed that these values were higher under treatment at 0 °C, but lower under the treatment at 10 °C. Moreover, the values of the width of the first spreading leaf blade (x₇), width of second spreading leaf blade (x₉), and number of secondary radicles (x₁₅) were lower values under treatment at 0 °C, but gradually increased with increasing treatment temperature and then decreased at 10 °C, while the values of plant height (x₁), stem base diameter (x₂), coleoptile length (x₃), leaf sheath length of first spreading leaf (x₄), leaf sheath length of second spreading leaf (x₅), length of first spreading leaf blade (x₆), length of second spreading leaf blade (x₈), length of third spreading leaf blade (x₁₀), width of third spreading leaf blade (x₁₁), total leaf area of first to third spreading leaves (x₁₂), primordial radicle length (x₁₃), root volume (x₁₇), aboveground fresh weight (x₂₂), belowground fresh weight (x₂₃), aboveground dry weight (x₂₅), and belowground dry weight (x₂₆) were lower at 0 °C, and were then progressively increased with increasing temperature treatment, reaching their maximum values at 10 °C.

3.1.2. Effect of Different Stages of Low-Temperature Treatment and Duration on Maize Yield Per Plant from Sowing to Seedling Stage

IBM SPSS Statistics 19 (IBM Corp., Armonk, NY, USA) was used to perform multifactorial analysis of variance (ANOVA) to analyze the significance of differences between different treatment combinations. We calculated the interaction effect between factors of the two treatments using Duncan’s method of multiple comparisons. We observed highly significant effects on maize yield per plant at different treatment stages under different low-temperature and treatment durations (p < 0.01). Furthermore, we observed a highly significant interaction effect (p < 0.01) between the effects of different treatment temperatures and low-temperature treatment durations on maize yield per plant (Figure 4).

We found that the maize yield per plant showed a gradual decrease with the decrease in treatment temperature and prolongation of low-temperature treatment (Figure 4). At the P1 stage, as low-temperature treatment times were extended and treatment temperatures lowered, the maize yield per plant was gradually decreased, reaching the lowest value under 0 °C treatment for 5 d, which was 67.22% lower than that of the maize yield per plant under no low-temperature treatment (control); conversely, under 10 °C treatment for 1 d, the yield per plant reached the maximum value, which was 11.55% higher than that of the control condition. In the P2 and P3 stages, all maize seedlings died at 0 °C, and the maize yield per plant was the lowest under 5 d treatment at 2 °C, at 79.46 g and 88.04 g, respectively, exhibiting decreases of 68.58% and 65.19% compared with that of the control group. At the P2 and P3 stages, all maize seedlings died at 0 °C, and the maize yield per plant was the lowest under 5 d treatment at 2 °C, at 79.46 g and 88.04 g, respectively, which were 68.58% and 65.19% lower than that of the control, respectively. Yield per plant was the highest under 1 d treatment at 10 °C for P2 and P3 stages, at 283.23 g and 280.37 g, respectively, which were 12.00% and 10.87% higher than that of the control. At the P4, P5, and P6 stages, all maize seedlings under the 0 °C treatment died, including all seedlings treated at 2 °C for 2, 3, 4, and 5 d at the P4 stage, and all seedlings treated at 2 °C for 3, 4, and 5 d at the P5 and P6 stages. At the P4, P5, and P6 stages, the maize yield per plant was the lowest under 5 d of 4 °C treatment, with 36.08%, 47.30%, and 44.55% lower yield per plant compared with that of maize grown at the control temperature, whereas the maize yield per plant was highest under 1 d of 10 °C treatment, with 7.07%, 2.36%, and 10.32% higher yield per plant compared with that of the control.

3.1.3. Classification of Seedling Quality

We determined that the maize yield per plant ranged from 80.56 to 287.19 g (Figure 5). This range provided a comprehensive coverage of the yield per plant of the test varieties at different stages from sowing to seedling under the influence of different temperatures and low-temperature treatment durations. Using cluster analysis of the yield per plant under each treatment combination [54,55,56], we classified the corresponding maize seedling quality into four grades, with the yield per plant in the range of 225.82–287.19 g as the optimal seedling grade (I), range of 186.69–225.82 g as the sub-optimal seedling grade (II), range of 118.93–186.69 g as the medium seedling grade (III), and range of 80.56–118.93 g as the weak seedling grade (IV).

3.2. Analysis of Seedling Quality Grading Model

3.2.1. Training Results and Analysis

The quality indicators of 3623 seedlings of ‘Xianyu335’ were divided into training and testing sets at a 3:1 ratio. The training set consisted of 2718 seedling quality evaluation indicators, including 918 optimal seedlings, 864 sub-optimal seedlings, 702 medium seedlings, and 234 weak seedlings. However, the test set comprised 305 optimal seedlings, 288 sub-optimal seedlings, 234 medium seedlings, and 78 weak seedlings, with a total of 905 seedling quality evaluation index datasets.

The sample dataset of this study was 3623 × 27 (3623 seedling plants were obtained, each involving 27 seedling phenotypic agronomic traits), and the training and test sets were divided randomly according to the 3:1 ratio. We adopted the exclusive hot coding to add labels to sample objects, and the generated classification labels are presented in

Table 1. Sample coding and quantitative division of maize seedling quality in different classes.

Seedling Quality Class	Number of Training Set	Number of Tests	Coded Label
Optimal seedling	918	305	[1, 0, 0, 0]
Sub-optimal seedling	864	288	[0, 1, 0, 0]
Medium seedling	702	234	[0, 0, 1, 0]
Weak seedling	234	78	[0, 0, 0, 1]
Total	2718	905	—

.

We fed the training set samples into the CNN model and trained the network parameters using Equations (5)–(10) of the Section 2.3.2 optimizer. We also set the number of iterations of the network model epoch to 100, the batch size to 16, and the convolution kernel size to 5, 7, and 6. When the training accuracy reached its highest value at a minimum loss, the model performed best. The training set accuracy and loss transformation curve are shown in Figure 6.

As shown in Figure 6a, the accuracy curve of the sample training set of the model was relatively smooth. We observed that when the number of training reached 84, the accuracy reached its maximum value of 98.16%. The training time of the model was 14 s, with the accuracy curve transforming in the range of 0.6259–1. As shown in Figure 6b, the training set loss value curve of the model varied from 0.8503 to 0.0007, suggesting good prediction performance of the model. These findings indicated that by combining seedling phenotypic indices and a one-dimensional CNN as proposed in this study, the maize seedling quality classes can be effectively identified.

3.2.2. Test Results and Analysis

We used a one-dimensional convolutional neural network (CNN) to construct a model for grading the quality of cold region maize seedlings in this study and inputted 27 seedling quality indicators into the trained model to evaluate its performance. The process of generating the maize seedling quality grading model was shown in Figure 7:

During the process of simulation, the decoding mapping rule of the actual output value of the LeNet-5 network corresponding to the exclusive hot coding of maize seedling quality type was as follows: when the network output was $y_1=y_{\max }\left(y_1,y_2,y_3,y_4\right)$, the corresponding optimal seedling was encoded as [1, 0, 0, 0]; when the network output was $y_2=y_{\max }\left(y_1,y_2,y_3,y_4\right)$, the corresponding sub-optimal seedling was encoded as [0, 1, 0, 0]; when the network output was $y_3=y_{\max }\left(y_1,y_2,y_3,y_4\right)$, the corresponding medium seedling was encoded as [0, 0, 1, 0]; when the network output was $y_4=y_{\max }\left(y_1,y_2,y_3,y_4\right)$, the corresponding weak seedling was encoded as [0, 0, 0, 1]. Following the above rules, the network output vector could be parsed to the output level of maize seedlings.

The calculated values of the LeNet-5 network model are shown in Figure 8. Overall, we used 905 sets of phenotypic data of four kinds of maize seedlings as the test set of the model to verify the trained classification model of maize seedlings in cold regions. The four colors represent the four types of maize seedlings, that is, the four columns of node values that the network actually outputs, with only the correctly graded seedlings in each individual column of nodes having the largest actual output value from the network. The horizontal coordinate $x\in \left[1,305\right]$ in Figure 8a corresponds to the optimal seedling for maize, with the average error of the simulation test being 0.0111. As seen in Figure 8a, the optimal seedling was misclassified and incorrectly recognized as a sub-optimal seedling at x = 21, 58, 170, 188, 190, and 234; The horizontal coordinate $x\in \left[306,593\right]$ in Figure 8b corresponds to the sub-optimal seedling for maize, with the average error of the simulation test being 0.0161, with a total of 14 recognition errors; The horizontal coordinate $x\in \left[594,827\right]$ in Figure 8c corresponds to the medium seedling for maize, with the simulation test showing 7 recognition errors at an average error of 0.0118. The horizontal coordinate $x\in \left[828,905\right]$ in Figure 8d corresponds to the weak seedling, all of which were identified. The recognition accuracy of the LeNet-5 network model was 97.23%, achieving a good recognition effect, and enabling the quick and efficient realization of the classification of maize seedlings.

3.2.3. Model Evaluation

To evaluate the performance of the maize seedling quality classification model in this study, we selected the precision, recall, and F1-Score as the evaluation indices of the network model, calculated the indices of the training and test sets in several common models based on Formulas (11)–(14), and evaluated the performance of the prediction effect.

According to

Table 2. Performance evaluation of maize seedling quality classification model.

Evaluation Indicators Type of Seedlings	Training Set			Test Set
	Precision	Recall	F1-Score	Precision	Recall	F1-Score
Optimal seedling	99.0%	99.3%	99.1%	98.0%	99.0%	98.5%
Sub-optimal seedling	98.4%	98.2%	98.3%	97.9%	97.2%	97.5%
Medium seedling	98.6%	98.2%	98.4%	96.6%	98.7%	97.6%
Weak seedling	98.3%	99.1%	98.7%	100.0%	92.9%	96.3%

, the prediction of the training set of the classification model was above 98%, with the prediction of the test set being above 96%. We found that the highest and lowest recalls were 99.3% and 98.2% for the training set and 99.0% and 92.9% for the test set, respectively. The average F1-Score for the training set of the model was 98.63%, while the average F1-Score for the test set was 97.48%. The difference between the maximum and minimum precision, recall, and F1-Score of the LeNet-5 model was 3.4%, 6.4%, and 3.4%, respectively, suggesting the good performance of the model.

Given the possible similarities and differences between maize samples corresponding to different seedling quality grades, we established a confusion matrix. This was used to determine whether the model constructed in this study could identify different grades of maize seedlings. Figure 9 shows the confusion matrix for the sample test set of maize seedling quality indicators, with the horizontal coordinates representing the sample predicted labels, vertical coordinates representing the sample true labels, and diagonal representing the accuracy of identifying samples. As shown in Figure 9, the weak seedlings in the classification of maize seedling quality were completely identified, with the test set samples classified as optimal, sub-optimal, and medium seedlings showing a prediction error of 0.02, 0.06, and 0.04, respectively. This identification confusion was attributed to the error of manually collecting phenotypic indices and to differences in the characteristics of different seedling individuals, which led to confusion in the prediction of test set samples; however, the overall prediction effect of the model was good.

We conducted tests and evaluations of the LeNet-5 grading performance using several common classification models: the backpropagation neural network (BP) [57] and partial least squares (PLS) [58] for grading prediction of seedling quality. As shown in Figure 10, the recognition accuracy of the LeNet-5 CNN model in this study was higher than that of the traditional 2 classification models in both the training and test sets. To observe the prediction effect of each model more intuitively, we compared the prediction results with those of the one-dimensional CNN-based model proposed in this study (

Table 3. Comparison of prediction results of different models.

Model	Training Set		Test Set
	Accuracy	Training Time/s	Accuracy	Simulation Time/s
BP	93.97%	5	95.80%	0.0033
LeNet5	98.16%	14	97.23%	0.0099
PLS	90.06%	3	89.89%	0.0022

).

As shown in Table 3, the LeNet-5 model had an accuracy of 98.16% in the training set and 97.23% in the test set. However, in the conventional classification model, the highest accuracy was 93.97% and 95.80% for the training and test sets, respectively, whereas the lowest accuracy was 90.06% and 89.89%, respectively. In addition, we observed that the training time of the LeNet-5 network was no more than 14 s, whereas that of the testing time was only 0.0099 s. The recognition time was shorter compared with that of the traditional model. Hence, we concluded that the proposed LeNet-5 network can identify the quality of cold region maize seedlings more quickly and accurately, outperforming other traditional machine learning models, and thus is an effective identification method, which can serve as a new reference and technical guide for constructing a classification model for maize seedling qualities using one-dimensional CNNs.

4. Discussion

In this study, we established a deep learning-based maize seedling quality grading model by extracting maize seedling phenotypic index parameters and achieved better experimental results regarding recognition accuracy and model efficiency. This method of seedling quality grading based on the fusion of phenotypic parameter extraction and the LeNet-5 model exhibited characteristics of high universality and adaptability and can provide a theoretical basis for the precise management of maize seedlings or even other crops in the field.

4.1. Analysis of the Methodological Strengths of This Study

The excellent seedling quality and neatness of maize crops are major factors in determining the abundance, quality, and efficiency of maize production [39,59]. To explore simple and convenient methods for evaluating the quality of maize, many researchers have assessed plant growth by collecting maize phenotypic indicators. Wang et al. [60] measured aboveground dry weight, underground dry weight, root length, and other indicators to evaluate maize seedling quality. Mustamu et al. [61] selected indicators such as root length, plant height, root crown ratio, aboveground fresh weight, underground fresh weight, aboveground dry weight, and underground dry weight based on traditional methods to evaluate maize seedling quality. Compared with previous studies, in the field of crop seedling quality, in the present study, 27 maize seedling phenotypic quality indicators were extracted by combining the growth characteristics of maize in the field, further enriching the comprehensive evaluation index system of maize seedling quality. In contrast, the current research reports on grading prediction of crop seedling quality are mostly based on image recognition to obtain the features of samples for analysis and evaluation [62,63,64], which not only requires expensive instruments but also uses a single evaluation index, establishing a large model with a long training time. In the present study, using the extracted phenotypic quality indicators of maize seedlings, we optimized the traditional agricultural model building method and established a deep learning-based maize seedling quality grading model, with a training accuracy of 98.16%, training time of 14 s, and training loss value of 0.012, which achieved better results in terms of recognition accuracy and model efficiency. Meanwhile, in this study, we replaced the soft-638 max function in the LeNet-5 model with the linear function and modified the LeNet-5 639 model to a regression CNN. The training accuracy of the regression model was 36.20%, 640 which was 61.96% lower than that of the LeNet-5 model. Perugachi-Diaz et al. [35] used CNN to predict the growth status of white cabbage seedlings, and the Alex Net model was used to accurately classify 94% of seedlings. Compared with the Alex Net model used by Perugachi-Diaz et al., the grading accuracy of this model was improved by 4.16%. Although Yuan et al. [65] used a type of point-centered CNN combined with embedded feature selection (PCNN-FS) to conduct non-destructive identification of moldy peanuts, with a training time of 41.88 s, the training time of the model in the present study was further reduced by 27.88 s compared with that of Yuan et al. [65]. Therefore, the model constructed in the present study achieves extremely high detection accuracy while reducing training time, making it an accurate, fast, and efficient maize seedling grading model.

4.2. Error Analysis

In the simulation of grading different maize seedlings in this study, the test sets of optimal, sub-optimal, and medium seedlings showed identification confusions of 0.02, 0.06, and 0.04, respectively, hence generating some errors compared with the actual conditions. This may be because of the idiosyncratic differences between different individual maize seedlings, especially the presence of singular samples affecting the high degree of consistency of phenotypic parameters. Another reason may be human subjective measurement bias during the data collection process, which affects the accuracy of the recognition model. To address this problem, researchers can use high-precision measuring instruments during data collection to avoid human error. Furthermore, the model needs to be optimized to adapt to specific research objectives, and the CNN learning algorithm should be optimized to improve model recognition accuracy and avoid model overfitting. Hence, by selecting the preferred phenotypic parameters and intelligent algorithms, the reliability and authenticity of the quality grading of maize seedlings would be improved.

4.3. Future Promotion and Application

In this study, we constructed a maize seedling grading model based on the LeNet-5 network, which can grade maize seedlings by inputting 27-dimensional seedling quality indices. We thus provided a new method and approach for rapid and accurate field grading of maize seedlings. Our model exhibited good robustness and fast computational speed. The simulation time for a single sample is only 0.0099 s, and the accuracy is as high as 97.23%. Future studies will aim to further develop and transplant the model to portable mobile terminals or devices [66], as well as collect maize seedling indicators from different regions and varieties and input them into the model for training to continuously improve the model [32,67] and enhance its universal applicability. Finally, it will be applied to specific production practices to help agricultural producers, researchers, and technological personnel achieve efficient evaluation and grading of maize seedlings quality under field conditions and quickly and accurately evaluate maize seedling quality.

5. Conclusions

The quality grading model developed in this study for cold region maize seedlings can comprehensively summarize the quality indices of maize during seedling formation with good stability and is simple and convenient to use. Seedling quality evaluation indices included plant height, total leaf area, radicle length, plant dry weight, and other important quality indices, which can comprehensively reflect the coordination between horizontal and vertical growth of maize plants, aboveground and belowground, fully embody the interdependent and mutual constraints of the growth and development of maize in cold areas, and comprehensively summarize the overall quality of seedlings. ‘Xianyu335’, which is widely used in the high-latitude cold region of Northeast China, was selected as the research object, and a combination of manual monitoring of maize agronomic index traits and CNN was used to establish a seedling quality grading model for maize seedlings in cold regions. We input a total of 27 seedling quality indices into the LeNet-5 network model. The results showed that the LeNet-5 model achieved 98.16% accuracy in grading the quality of maize seedlings, with an improvement of 4.19% in recognition accuracy compared with that of the BP and an improvement of 8.1% compared with that of the linear regression PLS. Therefore, the LeNet-5 model developed in this study is an efficient grading method for evaluating the quality of maize seedlings in cold areas and can quantify and quickly assess the quality of maize seedlings in actual production. Our study provides a theoretical and reference basis for realizing rapid and accurate evaluation of the quality of maize seedlings, carrying out timely interventions to control maize pests and diseases in cold regions, and promoting standardized and refined production management of maize.